2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[SCCfinal]{Modify and collect code generation for final STG program}
6 This is now a sort-of-normal STG-to-STG pass (WDP 94/06), run by stg2stg.
8 * Traverses the STG program collecting the cost centres. These are
9 required to declare the cost centres at the start of code
12 Note: because of cross-module unfolding, some of these cost centres
13 may be from other modules. But will still have to give them
14 "extern" declarations.
16 * Puts on CAF cost-centres if the user has asked for individual CAF
19 * Ditto for individual DICT cost-centres.
21 * Boxes top-level inherited functions passed as arguments.
23 * "Distributes" given cost-centres to all as-yet-unmarked RHSs.
26 module SCCfinal ( stgMassageForProfiling ) where
28 #include "HsVersions.h"
32 import CmdLineOpts ( opt_AutoSccsOnIndividualCafs )
33 import CostCentre -- lots of things
34 import Id ( Id, mkSysLocal, idType, idName )
35 import Module ( Module )
36 import UniqSupply ( uniqFromSupply, splitUniqSupply, UniqSupply )
37 import Unique ( Unique )
38 import Type ( splitForAllTys, splitTyConApp_maybe )
39 import TyCon ( isFunTyCon )
42 import Name ( isLocallyDefined )
43 import Util ( removeDups )
46 infixr 9 `thenMM`, `thenMM_`
50 type CollectedCCs = ([CostCentre], -- locally defined ones
51 [CostCentre], -- ones needing "extern" decls
52 [CostCentreStack]) -- singleton stacks (for CAFs)
54 stgMassageForProfiling
55 :: Module -- module name
56 -> UniqSupply -- unique supply
57 -> [StgBinding] -- input
58 -> (CollectedCCs, [StgBinding])
60 stgMassageForProfiling mod_name us stg_binds
62 ((local_ccs, extern_ccs, cc_stacks),
64 = initMM mod_name us (do_top_bindings stg_binds)
66 (fixed_ccs, fixed_cc_stacks)
67 = if opt_AutoSccsOnIndividualCafs
68 then ([],[]) -- don't need "all CAFs" CC
69 -- (for Prelude, we use PreludeCC)
70 else ([all_cafs_cc], [all_cafs_ccs])
72 local_ccs_no_dups = fst (removeDups cmpCostCentre local_ccs)
73 extern_ccs_no_dups = fst (removeDups cmpCostCentre extern_ccs)
75 ((fixed_ccs ++ local_ccs_no_dups,
77 fixed_cc_stacks ++ cc_stacks), stg_binds2)
80 all_cafs_cc = mkAllCafsCC mod_name
81 all_cafs_ccs = mkSingletonCCS all_cafs_cc
84 do_top_bindings :: [StgBinding] -> MassageM [StgBinding]
86 do_top_bindings [] = returnMM []
88 do_top_bindings (StgNonRec b rhs : bs)
89 = do_top_rhs b rhs `thenMM` \ rhs' ->
91 do_top_bindings bs `thenMM` \bs' ->
92 returnMM (StgNonRec b rhs' : bs')
95 do_top_bindings (StgRec pairs : bs)
96 = addTopLevelIshIds binders (
97 mapMM do_pair pairs `thenMM` \ pairs2 ->
98 do_top_bindings bs `thenMM` \ bs' ->
99 returnMM (StgRec pairs2 : bs')
102 binders = map fst pairs
104 = do_top_rhs b rhs `thenMM` \ rhs2 ->
108 do_top_rhs :: Id -> StgRhs -> MassageM StgRhs
110 do_top_rhs binder (StgRhsClosure _ bi srt fv u [] (StgSCC cc (StgConApp con args)))
111 | not (isSccCountCostCentre cc) && not (isDllConApp con args)
112 -- Trivial _scc_ around nothing but static data
113 -- Eliminate _scc_ ... and turn into StgRhsCon
115 -- isDllConApp checks for LitLit args too
116 = returnMM (StgRhsCon dontCareCCS con args)
118 {- Can't do this one with cost-centre stacks: --SDM
119 do_top_rhs binder (StgRhsClosure no_cc bi srt fv u [] (StgSCC ty cc expr))
120 | (noCCSAttached no_cc || currentOrSubsumedCCS no_cc)
121 && not (isSccCountCostCentre cc)
122 -- Top level CAF without a cost centre attached
123 -- Attach and collect cc of trivial _scc_ in body
124 = collectCC cc `thenMM_`
125 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
126 returnMM (StgRhsClosure cc bi srt fv u [] expr')
129 do_top_rhs binder (StgRhsClosure no_cc bi srt fv u [] body)
130 | noCCSAttached no_cc || currentOrSubsumedCCS no_cc
131 -- Top level CAF without a cost centre attached
132 -- Attach CAF cc (collect if individual CAF ccs)
133 = (if opt_AutoSccsOnIndividualCafs
134 then let cc = mkAutoCC binder mod_name CafCC
135 ccs = mkSingletonCCS cc
137 collectCC cc `thenMM_`
138 collectCCS ccs `thenMM_`
141 returnMM all_cafs_ccs) `thenMM` \ caf_ccs ->
142 set_prevailing_cc caf_ccs (do_expr body) `thenMM` \ body' ->
143 returnMM (StgRhsClosure caf_ccs bi srt fv u [] body')
145 do_top_rhs binder (StgRhsClosure cc bi srt fv u [] body)
146 -- Top level CAF with cost centre attached
147 -- Should this be a CAF cc ??? Does this ever occur ???
148 = pprPanic "SCCfinal: CAF with cc:" (ppr cc)
150 {- can't do this with cost-centre stacks: --SDM
151 do_top_rhs binder (StgRhsClosure _ bi srt fv u args (StgSCC cc expr))
152 | not (isSccCountCostCentre cc)
153 -- Top level function with trivial _scc_ in body
154 -- Attach and collect cc of trivial _scc_
155 = collectCC cc `thenMM_`
156 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
157 returnMM (StgRhsClosure cc bi srt fv u args expr')
160 do_top_rhs binder (StgRhsClosure no_ccs bi srt fv u args body)
161 -- Top level function, probably subsumed
162 | noCCSAttached no_ccs
163 = set_lambda_cc (do_expr body) `thenMM` \ body' ->
164 returnMM (StgRhsClosure subsumedCCS bi srt fv u args body')
167 = pprPanic "SCCfinal: CAF with cc:" (ppr no_ccs)
169 do_top_rhs binder (StgRhsCon ccs con args)
170 -- Top-level (static) data is not counted in heap
171 -- profiles; nor do we set CCCS from it; so we
172 -- just slam in dontCareCostCentre
173 = returnMM (StgRhsCon dontCareCCS con args)
176 do_expr :: StgExpr -> MassageM StgExpr
178 do_expr (StgLit l) = returnMM (StgLit l)
180 do_expr (StgApp fn args)
181 = boxHigherOrderArgs (StgApp fn) args
183 do_expr (StgConApp con args)
184 = boxHigherOrderArgs (\args -> StgConApp con args) args
186 do_expr (StgPrimApp con args res_ty)
187 = boxHigherOrderArgs (\args -> StgPrimApp con args res_ty) args
189 do_expr (StgSCC cc expr) -- Ha, we found a cost centre!
190 = collectCC cc `thenMM_`
191 do_expr expr `thenMM` \ expr' ->
192 returnMM (StgSCC cc expr')
194 do_expr (StgCase expr fv1 fv2 bndr srt alts)
195 = do_expr expr `thenMM` \ expr' ->
196 do_alts alts `thenMM` \ alts' ->
197 returnMM (StgCase expr' fv1 fv2 bndr srt alts')
199 do_alts (StgAlgAlts ty alts def)
200 = mapMM do_alt alts `thenMM` \ alts' ->
201 do_deflt def `thenMM` \ def' ->
202 returnMM (StgAlgAlts ty alts' def')
204 do_alt (id, bs, use_mask, e)
205 = do_expr e `thenMM` \ e' ->
206 returnMM (id, bs, use_mask, e')
208 do_alts (StgPrimAlts ty alts def)
209 = mapMM do_alt alts `thenMM` \ alts' ->
210 do_deflt def `thenMM` \ def' ->
211 returnMM (StgPrimAlts ty alts' def')
214 = do_expr e `thenMM` \ e' ->
217 do_deflt StgNoDefault = returnMM StgNoDefault
218 do_deflt (StgBindDefault e)
219 = do_expr e `thenMM` \ e' ->
220 returnMM (StgBindDefault e')
223 = do_let b e `thenMM` \ (b,e) ->
224 returnMM (StgLet b e)
226 do_expr (StgLetNoEscape lvs1 lvs2 b e)
227 = do_let b e `thenMM` \ (b,e) ->
228 returnMM (StgLetNoEscape lvs1 lvs2 b e)
230 ----------------------------------
232 do_let (StgNonRec b rhs) e
233 = do_rhs rhs `thenMM` \ rhs' ->
235 do_expr e `thenMM` \ e' ->
236 returnMM (StgNonRec b rhs',e')
239 do_let (StgRec pairs) e
240 = addTopLevelIshIds binders (
241 mapMM do_pair pairs `thenMM` \ pairs' ->
242 do_expr e `thenMM` \ e' ->
243 returnMM (StgRec pairs', e')
246 binders = map fst pairs
248 = do_rhs rhs `thenMM` \ rhs2 ->
251 ----------------------------------
252 do_rhs :: StgRhs -> MassageM StgRhs
253 -- We play much the same game as we did in do_top_rhs above;
254 -- but we don't have to worry about cafs etc.
257 do_rhs (StgRhsClosure closure_cc bi srt fv u [] (StgSCC ty cc (StgCon (DataCon con) args _)))
258 | not (isSccCountCostCentre cc)
259 = collectCC cc `thenMM_`
260 returnMM (StgRhsCon cc con args)
264 do_rhs (StgRhsClosure _ bi srt fv u args (StgSCC ty cc expr))
265 | not (isSccCountCostCentre cc)
266 = collectCC cc `thenMM_`
267 set_prevailing_cc cc (do_expr expr) `thenMM` \ expr' ->
268 returnMM (StgRhsClosure cc bi srt fv u args expr')
271 do_rhs (StgRhsClosure cc bi srt fv u [] body)
272 = do_expr body `thenMM` \ body' ->
273 returnMM (StgRhsClosure currentCCS bi srt fv u [] body')
275 do_rhs (StgRhsClosure cc bi srt fv u args body)
276 = set_lambda_cc (do_expr body) `thenMM` \ body' ->
277 get_prevailing_cc `thenMM` \ prev_ccs ->
278 returnMM (StgRhsClosure currentCCS bi srt fv u args body')
280 do_rhs (StgRhsCon cc con args)
281 = returnMM (StgRhsCon currentCCS con args)
284 %************************************************************************
286 \subsection{Boxing higher-order args}
288 %************************************************************************
290 Boxing is *turned off* at the moment, until we can figure out how to
291 do it properly in general.
295 :: ([StgArg] -> StgExpr)
296 -- An application lacking its arguments
297 -> [StgArg] -- arguments which we might box
300 #ifndef PROF_DO_BOXING
301 boxHigherOrderArgs almost_expr args
302 = returnMM (almost_expr args)
304 boxHigherOrderArgs almost_expr args
305 = getTopLevelIshIds `thenMM` \ ids ->
306 mapAccumMM (do_arg ids) [] args `thenMM` \ (let_bindings, new_args) ->
307 returnMM (foldr (mk_stg_let currentCCS) (almost_expr new_args) let_bindings)
311 do_arg ids bindings arg@(StgVarArg old_var)
312 | (not (isLocallyDefined old_var) || elemVarSet old_var ids)
313 && isFunType var_type
314 = -- make a trivial let-binding for the top-level function
315 getUniqueMM `thenMM` \ uniq ->
317 new_var = mkSysLocal SLIT("sf") uniq var_type
319 returnMM ( (new_var, old_var) : bindings, StgVarArg new_var )
321 var_type = idType old_var
323 do_arg ids bindings arg = returnMM (bindings, arg)
326 mk_stg_let :: CostCentreStack -> (Id, Id) -> StgExpr -> StgExpr
328 mk_stg_let cc (new_var, old_var) body
330 rhs_body = StgApp old_var [{-args-}]
331 rhs_closure = StgRhsClosure cc stgArgOcc NoSRT [{-fvs-}] ReEntrant [{-args-}] rhs_body
333 StgLet (StgNonRec new_var rhs_closure) body
335 bOGUS_LVs = emptyUniqSet -- easier to print than: panic "mk_stg_let: LVs"
338 = case splitForAllTys var_type of
339 (_, ty) -> case splitTyConApp_maybe ty of
340 Just (tycon,_) | isFunTyCon tycon -> True
345 %************************************************************************
347 \subsection{Boring monad stuff for this}
349 %************************************************************************
353 = Module -- module name
354 -> CostCentreStack -- prevailing CostCentre
355 -- if none, subsumedCosts at top-level
356 -- useCurrentCostCentre at nested levels
358 -> VarSet -- toplevel-ish Ids for boxing
360 -> (CollectedCCs, result)
362 -- the initMM function also returns the final CollectedCCs
364 initMM :: Module -- module name, which we may consult
369 initMM mod_name init_us m = m mod_name noCCS init_us emptyVarSet ([],[],[])
371 thenMM :: MassageM a -> (a -> MassageM b) -> MassageM b
372 thenMM_ :: MassageM a -> (MassageM b) -> MassageM b
374 thenMM expr cont mod scope_cc us ids ccs
375 = case splitUniqSupply us of { (s1, s2) ->
376 case (expr mod scope_cc s1 ids ccs) of { (ccs2, result) ->
377 cont result mod scope_cc s2 ids ccs2 }}
379 thenMM_ expr cont mod scope_cc us ids ccs
380 = case splitUniqSupply us of { (s1, s2) ->
381 case (expr mod scope_cc s1 ids ccs) of { (ccs2, _) ->
382 cont mod scope_cc s2 ids ccs2 }}
384 returnMM :: a -> MassageM a
385 returnMM result mod scope_cc us ids ccs = (ccs, result)
388 nopMM mod scope_cc us ids ccs = (ccs, ())
390 mapMM :: (a -> MassageM b) -> [a] -> MassageM [b]
391 mapMM f [] = returnMM []
393 = f m `thenMM` \ r ->
394 mapMM f ms `thenMM` \ rs ->
397 mapAccumMM :: (acc -> x -> MassageM (acc, y)) -> acc -> [x] -> MassageM (acc, [y])
398 mapAccumMM f b [] = returnMM (b, [])
399 mapAccumMM f b (m:ms)
400 = f b m `thenMM` \ (b2, r) ->
401 mapAccumMM f b2 ms `thenMM` \ (b3, rs) ->
404 getUniqueMM :: MassageM Unique
405 getUniqueMM mod scope_cc us ids ccs = (ccs, uniqFromSupply us)
407 addTopLevelIshId :: Id -> MassageM a -> MassageM a
408 addTopLevelIshId id scope mod scope_cc us ids ccs
409 | isCurrentCCS scope_cc = scope mod scope_cc us ids ccs
410 | otherwise = scope mod scope_cc us (extendVarSet ids id) ccs
412 addTopLevelIshIds :: [Id] -> MassageM a -> MassageM a
413 addTopLevelIshIds [] cont = cont
414 addTopLevelIshIds (id:ids) cont
415 = addTopLevelIshId id (addTopLevelIshIds ids cont)
417 getTopLevelIshIds :: MassageM VarSet
418 getTopLevelIshIds mod scope_cc us ids ccs = (ccs, ids)
421 The prevailing CCS is used to tell whether we're in a top-levelish
422 position, where top-levelish is defined as "not inside a lambda".
423 Prevailing CCs used to be used for something much more complicated,
427 set_lambda_cc :: MassageM a -> MassageM a
428 set_lambda_cc action mod scope_cc us ids ccs
429 = action mod currentCCS us ids ccs
431 set_prevailing_cc :: CostCentreStack -> MassageM a -> MassageM a
432 set_prevailing_cc cc_to_set_to action mod scope_cc us ids ccs
433 = action mod cc_to_set_to us ids ccs
435 get_prevailing_cc :: MassageM CostCentreStack
436 get_prevailing_cc mod scope_cc us ids ccs = (ccs, scope_cc)
440 collectCC :: CostCentre -> MassageM ()
442 collectCC cc mod_name scope_cc us ids (local_ccs, extern_ccs, ccss)
443 = ASSERT(not (noCCAttached cc))
444 if (cc `ccFromThisModule` mod_name) then
445 ((cc : local_ccs, extern_ccs, ccss), ())
446 else -- must declare it "extern"
447 ((local_ccs, cc : extern_ccs, ccss), ())
449 collectCCS :: CostCentreStack -> MassageM ()
451 collectCCS ccs mod_name scope_cc us ids (local_ccs, extern_ccs, ccss)
452 = ASSERT(not (noCCSAttached ccs))
453 ((local_ccs, extern_ccs, ccs : ccss), ())